1 #include "parse_trigrams.h"
3 #include "unique_sort.h"
12 string print_td(const TrigramDisjunction &td)
14 if (td.read_trigrams.size() == 0) {
15 // Before we've done hash lookups (or none matched), so print all alternatives.
16 if (td.trigram_alternatives.size() == 1) {
17 return print_trigram(td.trigram_alternatives[0]);
22 for (uint32_t trgm : td.trigram_alternatives) {
25 ret += print_trigram(trgm);
31 // Print only those that we actually have in the index.
32 if (td.read_trigrams.size() == 1) {
33 return print_trigram(td.read_trigrams[0].first.trgm);
38 for (auto &[trgmptr, len] : td.read_trigrams) {
41 ret += print_trigram(trgmptr.trgm);
49 string print_trigram(uint32_t trgm)
52 char(trgm & 0xff), char((trgm >> 8) & 0xff), char((trgm >> 16) & 0xff)
56 for (unsigned i = 0; i < 3;) {
61 } else if (int(ch[i]) >= 32 && int(ch[i]) <= 127) { // Holds no matter whether char is signed or unsigned.
65 // See if we have an entire UTF-8 codepoint, and that it's reasonably printable.
66 mbtowc(nullptr, 0, 0);
68 int ret = mbtowc(&pwc, ch + i, 3 - i);
69 if (ret >= 1 && pwc >= 32) {
70 str.append(ch + i, ret);
74 snprintf(buf, sizeof(buf), "\\x{%02x}", (unsigned char)ch[i]);
84 pair<uint32_t, size_t> read_unigram(const string &s, size_t start)
86 if (start >= s.size()) {
87 return { PREMATURE_END_UNIGRAM, 0 };
89 if (s[start] == '\\') {
91 if (start + 1 >= s.size()) {
92 return { PREMATURE_END_UNIGRAM, 1 };
94 return { (unsigned char)s[start + 1], 2 };
97 if (s[start] == '*' || s[start] == '?') {
99 return { WILDCARD_UNIGRAM, 1 };
101 if (s[start] == '[') {
102 // Character class; search to find the end.
104 if (start + len >= s.size()) {
105 return { PREMATURE_END_UNIGRAM, len };
107 if (s[start + len] == '!') {
110 if (start + len >= s.size()) {
111 return { PREMATURE_END_UNIGRAM, len };
113 if (s[start + len] == ']') {
117 if (start + len >= s.size()) {
118 return { PREMATURE_END_UNIGRAM, len };
120 if (s[start + len] == ']') {
121 return { WILDCARD_UNIGRAM, len + 1 };
128 return { (unsigned char)s[start], 1 };
131 uint32_t read_trigram(const string &s, size_t start)
133 pair<uint32_t, size_t> u1 = read_unigram(s, start);
134 if (u1.first == WILDCARD_UNIGRAM || u1.first == PREMATURE_END_UNIGRAM) {
137 pair<uint32_t, size_t> u2 = read_unigram(s, start + u1.second);
138 if (u2.first == WILDCARD_UNIGRAM || u2.first == PREMATURE_END_UNIGRAM) {
141 pair<uint32_t, size_t> u3 = read_unigram(s, start + u1.second + u2.second);
142 if (u3.first == WILDCARD_UNIGRAM || u3.first == PREMATURE_END_UNIGRAM) {
145 return u1.first | (u2.first << 8) | (u3.first << 16);
148 struct TrigramState {
150 unsigned next_codepoint;
152 bool operator<(const TrigramState &other) const
154 if (next_codepoint != other.next_codepoint)
155 return next_codepoint < other.next_codepoint;
156 return buffered < other.buffered;
158 bool operator==(const TrigramState &other) const
160 return next_codepoint == other.next_codepoint &&
161 buffered == other.buffered;
165 void parse_trigrams_ignore_case(const string &needle, vector<TrigramDisjunction> *trigram_groups)
167 vector<vector<string>> alternatives_for_cp;
169 // Parse the needle into Unicode code points, and do inverse case folding
170 // on each to find legal alternatives. This is far from perfect (e.g. ß
171 // will not become ss), but it's generally the best we can do without
172 // involving ICU or the likes.
173 mbtowc(nullptr, 0, 0);
174 const char *ptr = needle.c_str();
175 unique_ptr<char[]> buf(new char[MB_CUR_MAX]);
176 while (*ptr != '\0') {
178 int ret = mbtowc(&ch, ptr, strlen(ptr));
185 alt.push_back(string(ptr, ret));
187 if (towlower(ch) != wint_t(ch)) {
188 ret = wctomb(buf.get(), towlower(ch));
189 alt.push_back(string(buf.get(), ret));
191 if (towupper(ch) != wint_t(ch) && towupper(ch) != towlower(ch)) {
192 ret = wctomb(buf.get(), towupper(ch));
193 alt.push_back(string(buf.get(), ret));
195 alternatives_for_cp.push_back(move(alt));
198 // Now generate all possible byte strings from those code points in order;
199 // e.g., from abc, we'd create a and A, then extend those to ab aB Ab AB,
200 // then abc abC aBc aBC and so on. Since we don't want to have 2^n
201 // (or even 3^n) strings, we only extend them far enough to cover at
202 // least three bytes; this will give us a set of candidate trigrams
203 // (the filename must have at least one of those), and then we can
204 // chop off the first byte, deduplicate states and continue extending
205 // and generating trigram sets.
207 // There are a few special cases, notably the dotted i (İ), where the
208 // UTF-8 versions of upper and lower case have different number of bytes.
209 // If this happens, we can have combinatorial explosion and get many more
210 // than the normal 8 states. We detect this and simply bomb out; it will
211 // never really happen in real strings, and stopping trigram generation
212 // really only means our pruning of candidates will be less effective.
213 vector<TrigramState> states;
214 states.push_back(TrigramState{ "", 0 });
217 // Extend every state so that it has buffered at least three bytes.
218 // If this isn't possible, we are done with the string (can generate
219 // no more trigrams).
220 bool need_another_pass;
222 need_another_pass = false;
223 vector<TrigramState> new_states;
224 for (const TrigramState &state : states) {
225 if (read_trigram(state.buffered, 0) != PREMATURE_END_UNIGRAM) {
226 // No need to extend this further.
227 new_states.push_back(state);
230 if (state.next_codepoint == alternatives_for_cp.size()) {
231 // We can't form a complete trigram from this alternative,
235 for (const string &rune : alternatives_for_cp[state.next_codepoint]) {
236 TrigramState new_state{ state.buffered + rune, state.next_codepoint + 1 };
237 if (read_trigram(state.buffered, 0) == PREMATURE_END_UNIGRAM) {
238 need_another_pass = true;
240 new_states.push_back(move(new_state));
243 states = move(new_states);
244 } while (need_another_pass);
246 // OK, so now we have a bunch of states, and all of them are at least
247 // three bytes long. This means we have a complete set of trigrams,
248 // and the destination filename must contain at least one of them.
249 // Output those trigrams, cut out the first byte and then deduplicate
250 // the states before we continue.
251 bool any_wildcard = false;
252 vector<uint32_t> trigram_alternatives;
253 for (TrigramState &state : states) {
254 trigram_alternatives.push_back(read_trigram(state.buffered, 0));
255 state.buffered.erase(0, read_unigram(state.buffered, 0).second);
256 assert(trigram_alternatives.back() != PREMATURE_END_UNIGRAM);
257 if (trigram_alternatives.back() == WILDCARD_UNIGRAM) {
258 // If any of the candidates are wildcards, we need to drop the entire OR group.
259 // (Most likely, all of them would be anyway.) We need to keep stripping out
260 // the first unigram from each state.
264 unique_sort(&trigram_alternatives); // Could have duplicates, although it's rare.
265 unique_sort(&states);
268 TrigramDisjunction new_pt;
269 new_pt.remaining_trigrams_to_read = trigram_alternatives.size();
270 new_pt.trigram_alternatives = move(trigram_alternatives);
271 new_pt.max_num_docids = 0;
272 trigram_groups->push_back(move(new_pt));
275 if (states.size() > 100) {
276 // A completely crazy pattern with lots of those special characters.
277 // We just give up; this isn't a realistic scenario anyway.
278 // We already have lots of trigrams that should reduce the amount of
285 void parse_trigrams(const string &needle, bool ignore_case, vector<TrigramDisjunction> *trigram_groups)
288 parse_trigrams_ignore_case(needle, trigram_groups);
292 // The case-sensitive case is straightforward.
293 for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
294 uint32_t trgm = read_trigram(needle, i);
295 if (trgm == WILDCARD_UNIGRAM || trgm == PREMATURE_END_UNIGRAM) {
296 // Invalid trigram, so skip.
300 TrigramDisjunction new_pt;
301 new_pt.remaining_trigrams_to_read = 1;
302 new_pt.trigram_alternatives.push_back(trgm);
303 new_pt.max_num_docids = 0;
304 trigram_groups->push_back(move(new_pt));